library(dplyr); library(tidyr); library(purrr) # Data wrangling
library(ggplot2); library(stringr) # Plotting
library(tidyfit) # Auto-ML modeling
The Boston house prices data set (MASS::Boston
) presents
a popular test case for regression algorithms. In this article, I assess
the relative performance of 15 different linear regression techniques
using tidyfit::regress
.
Let’s begin by loading the data, and splitting it into equally sized training and test sets:
data <- MASS::Boston
# For reproducibility
set.seed(123)
ix_tst <- sample(1:nrow(data), round(nrow(data)*0.5))
data_trn <- data[-ix_tst,]
data_tst <- data[ix_tst,]
as_tibble(data)
#> # A tibble: 506 × 14
#> crim zn indus chas nox rm age dis rad tax ptratio black
#> <dbl> <dbl> <dbl> <int> <dbl> <dbl> <dbl> <dbl> <int> <dbl> <dbl> <dbl>
#> 1 0.00632 18 2.31 0 0.538 6.58 65.2 4.09 1 296 15.3 397.
#> 2 0.0273 0 7.07 0 0.469 6.42 78.9 4.97 2 242 17.8 397.
#> 3 0.0273 0 7.07 0 0.469 7.18 61.1 4.97 2 242 17.8 393.
#> 4 0.0324 0 2.18 0 0.458 7.00 45.8 6.06 3 222 18.7 395.
#> 5 0.0690 0 2.18 0 0.458 7.15 54.2 6.06 3 222 18.7 397.
#> 6 0.0298 0 2.18 0 0.458 6.43 58.7 6.06 3 222 18.7 394.
#> 7 0.0883 12.5 7.87 0 0.524 6.01 66.6 5.56 5 311 15.2 396.
#> 8 0.145 12.5 7.87 0 0.524 6.17 96.1 5.95 5 311 15.2 397.
#> 9 0.211 12.5 7.87 0 0.524 5.63 100 6.08 5 311 15.2 387.
#> 10 0.170 12.5 7.87 0 0.524 6.00 85.9 6.59 5 311 15.2 387.
#> # ℹ 496 more rows
#> # ℹ 2 more variables: lstat <dbl>, medv <dbl>
The aim is to predict the median value of owner-occupied homes
('medv'
in the data set) using the remaining columns.1 We will
assess performance using \(R^2\), which
allows intuitive comparisons of in-sample and out-of-sample fit.
To establish a baseline, we fit an OLS regression model and examine
the test and training set R-squared. The predict
function
can be used to obtain predictions, and yardstick
is used to
assess performance. The regression is estimated using
tidyfit::regress
. tidyfit::m
is a generic
model wrapper that is capable of fitting a large number of different
algorithms:
Let’s create a helper function to assess performance. The helper
function wraps predict
and prepares in and out-of-sample
performance stats:
assess <- function(model_frame, data_trn, data_tst) {
oos <- model_frame %>%
predict(data_tst) %>%
group_by(model, .add = TRUE) %>%
yardstick::rsq_trad(truth, prediction) %>%
mutate(type = "Out-of-sample") %>%
arrange(desc(.estimate))
is <- model_frame %>%
predict(data_trn) %>%
group_by(model, .add = TRUE) %>%
yardstick::rsq_trad(truth, prediction) %>%
mutate(type = "In-sample")
return(bind_rows(oos, is))
}
plotter <- function(df) {
df %>%
mutate(lab = round(.estimate, 2)) %>%
mutate(model = str_wrap(model, 12)) %>%
mutate(model = factor(model, levels = unique(.$model))) %>%
ggplot(aes(model, .estimate)) +
geom_point(aes(color = type), size = 2.5, shape = 4) +
geom_label(aes(label = lab, color = type), size = 2, nudge_x = 0.35) +
theme_bw() +
scale_color_manual(values = c("firebrick", "darkblue")) +
theme(legend.title = element_blank(),
axis.title.x = element_blank()) +
coord_cartesian(ylim = c(0.65, 0.95))
}
Now we plot the resulting metrics:
The in-sample fit of the regression is naturally better than out-of-sample, but the difference is not large. The results can be improved substantially by adding interaction terms between features. This leads to a vast improvement in in-sample fit (again, this is a logical consequence of including a large number of regressors), but leads to no improvement in out-of-sample fit, suggesting substantial overfit:
Regularization controls overfit by introducing parameter bias in exchange for a reduction in estimator variance. A large number of approaches to regularization exist that can be summarized in at least 5 categories:
tidyfit
includes linear estimators from each of these
categories, as summarized below:
Best subset | L1/L2 penalized | Latent factors | Machine Learning | Bayesian |
---|---|---|---|---|
Exhaustive search (leaps )
|
Lasso (glmnet )
|
Principal components regression (pls )
|
Gradient Boosting (mboost )
|
Bayesian regression (arm )
|
Forward selection (leaps )
|
Ridge (glmnet )
|
Partial least squares (pls )
|
Hierarchical feature regression (hfr )
|
Bayesian model averaging (BMS )
|
Backward selection (leaps )
|
ElasticNet (glmnet )
|
Support vector regression (e1071 )
|
Bayesian Lasso (monomvn )
|
|
Sequential Replacement (leaps )
|
Adaptive Lasso (glmnet )
|
Bayesian Ridge (monomvn )
|
||
Genetic Algorithm (‘gaselect’) | Spike and Slab Regression (‘BoomSpikeSlab’) |
In the next code chunk, 17 of the available regression techniques are
fitted to the data set. Hyperparameters are optimized using a 10-fold
cross validation powered by rsample
in the background. As a
practical aside, it is possible to set a future
plan to
ensure hyperparameter tuning is performed on a parallelized backend.2 This is
done using plan(multisession(workers = n_workers))
, where
n_workers
is the number of available cores. Furthermore, a
progress bar could be activated using
progressr::handlers(global = TRUE)
.
model_frame_all <- data_trn %>%
regress(medv ~ .*.,
OLS = m("lm"),
BAYES = m("bayes"),
BMA = m("bma", iter = 10000),
SEQREP = m("subset", method = "seqrep", IC = "AIC"),
`FORWARD SELECTION` = m("subset", method = "forward", IC = "AIC"),
`BACKWARD SELECTION` = m("subset", method = "backward", IC = "AIC"),
LASSO = m("lasso"),
BLASSO = m("blasso"),
SPIKESLAB = m("spikeslab", niter = 10000),
RIDGE = m("ridge"),
BRIDGE = m("bridge"),
ELASTICNET = m("enet"),
ADALASSO = m("adalasso", lambda_ridge = c(0.001, 0.01, 0.1)),
PCR = m("pcr"),
PLSR = m("plsr"),
HFR = m("hfr"),
`GRADIENT BOOSTING` = m("boost"),
SVR = m("svm"),
GENETIC = m("genetic", populationSize = 1000, numGenerations = 50, statistic = "AIC", maxVariables = 20),
.cv = "vfold_cv", .cv_args = list(v = 10))
Let’s use yardstick
again to calculate in-sample and
out-of-sample performance statistics. The models are arranged descending
order of test set performance:
The Bayesian regression techniques overwhelmingly outperform the remaining methods, with the exception of the hierarchical feature regression, genetic algorithm and the adaptive lasso, both of which achieve relatively high out-of-sample performance. The subset selection algorithms achieve relatively good results, while the penalized and latent factor regressions perform worst. All methods improve on the OLS results.
tidyfit
makes it exceedingly simple to fit different
regularized linear regressions. The package ensures that the input and
output of the modeling engine, tidyfit::m
are standardized
and model results are comparable. For instance, whenever necessary,
features are standardized with the coefficients transformed back to the
original scale. Hyperparameter grids are set to reasonable starting
values and can be passed manually to tidyfit::m
.
The package further includes a tidyfit::classify
method
that assumes a binomial or multinomial response. Many of the methods
implemented can handle Gaussian and binomial responses. Furthermore, the
tidyfit::regress
and tidyfit::classify
methods
seamlessly integrate with grouped tibbles, making it extremely powerful
to estimate a large number of regressions for different data groupings.
Groups as well as response distributions are automatically respected by
the predict
method.
Have a look at ?MASS::Boston
for
definitions of the features.↩︎
Parallelizing happens at the level of cross validation
slices, so always consider if the overhead of parallel computation is
worth it. For instance, parallelizing will likely be very useful with
.cv = "loo_cv"
but can be less useful with
.cv = "vfold_cv"
.↩︎